Machine learning approaches to improve prostate cancer outcome.

Abstract

82 Background: High-risk Prostate Cancer (PCa) occurs in as much as 37% of PCa cases in England and Wales. This proportion is much higher than in other developed countries. Furthermore, there is an imprecision in the primary treatment of these patients. More specifically, in the administration of systemic therapy. Currently, surgeons prescribe radical prostatectomy and radiation, while radiation oncologists often combine radiation with androgen deprivation therapy (ADT). However, ADT has significant side effects and not all patients develop Biochemical Recurrence/Metastasis. Furthermore, clinical factors have limited ability to identify these patients. Consequently, we turn to biomarker discovery for more accurate disease management. Methods: We applied a machine learning strategy to identify discriminatory DNA Methylation patterns between patients with biochemical recurrence from those with stable disease in high-risk PCa (D’Amico Classification, T-stage ≤ 3a). Training and feature selection occurred in cross-validation (4 Folds, 10 reps) on 102 patients from the TCGA-PRAD dataset (26 BCR). Gene co-expression networks were derived from mRNA expression (|Pearson correlation| > 0.70). Network clustering maximised global modularity (Louvain algorithm) and unsupervised clustering evaluated complete linkage. Functional enrichment employed MiSigDB Hallmarks and g:ProfileR. Results: A linear support vector machine outperformed Naïve Bayes, Random Forest and Artificial Neural Networks with an AUC = 0.91 ± 0.14 (95% CI). This epigenomic test preferentially sub-stratifies high-risk patients, as it shows no improvement over the “No Information Rate” of intermediate- and low-risk patients (p-value > 0.05, t-test). We identified functional clusters from a co-expression network superimposed with the genes most correlated to the epigenomic test. The most enriched cluster reveals disturbances in MYC targets, cell cycle and RNA splicing. Conclusions: Preliminary results indicate that an epigenomic test identifies a sub-group of patients which could benefit from ADT at the stage of primary treatment, however it requires further validation.